Google

Google announced a series of AI-related model updates and products at Google's Cloud Next 2024, including Gemini 1.5 Pro, which offers native audio (speech) understanding for the first time, CodeGemma, a new model for code generation, Axion, the first in-house Arm processor, and more. (Heart of the Machine)

On April 7, the Financial Times reported that Google is considering charging for new premium features powered by generative artificial intelligence, in what would be the biggest change ever to Google's search business. For years Google has offered a free consumer service funded entirely by advertising, and the proposed changes to its search engine would mark the first time Google has put all of its core products behind a 'paywall'. According to three people with knowledge of Google's plans, options being considered by Google include adding certain artificial intelligence search features to its premium subscription service. One of the people with knowledge of the matter said engineers are developing what would be needed to deploy the service...

Microsoft and OpenAI are closely planning an ambitious data center project to build an AI supercomputer called Stargate, according to tech media outlet The Information, citing sources. The computer will be equipped with millions of dedicated server chips and is designed to power OpenAI's AI technology. The project is expected to cost up to $100 billion, which is reportedly 100 times the cost of some of the largest data centers today. Microsoft will underwrite the project's capital investment, showing...

Google researchers have published a paper introducing the ObjectDrop Bootstrapping Counterfactuals approach for realizing realistic-feeling object removal and insertion. Faced with the problem that diffusion models often generate images that violate the laws of physics, this method supports realistic object insertion, making effects such as occlusions, shadows, and reflections more realistic. Paper address: https://arxiv.org/pdf/2403.18818.pdf

The SELF-DISCOVER framework, developed by Google and the University of Southern California, enables large language models to discover and solve complex reasoning problems on their own. In multiple complex reasoning tests, the framework delivers up to 421 TP3T of performance improvement, significantly outperforming traditional chain-thinking methods.SELF-DISCOVER excels at tasks requiring world knowledge by integrating multiple reasoning modules, improving efficiency, and reflecting intrinsic task characteristics. (AIGC Open Community)

On Wednesday, local time, Google made an announcement pre-disclosing an AI feature that can help users generate travel itineraries and trip suggestions through natural language conversations. Google revealed that behind this AI itinerary feature, data from over 200 million global locations is covered, aggregating ideas from across the internet, as well as data from reviews, photos, business profile details, and other data submitted by users to Google. The feature is currently only available for internal beta users in the US to try out.

Google is testing an AI summary feature in Google search results. Even if users have not yet opted in to the Google SGE search generation experience lab feature, they can still see AI-summarized answers in search results.

The French market regulator announced on March 20 that it had issued a new fine of 250 million euros (about $272 million) to the U.S.-based Google for violating European Union intellectual property rights regulations by using the content of French publishers and news organizations to train the chatbot "Bard" (an upgraded version of which is called "Gemini") without consent. Gemini") without consent, in violation of European Union regulations on intellectual property rights.

Google has launched the TacticAI soccer tactics AI assistant. The system is able to provide tactical insights to experts through predictive and generative AI, and is particularly good at corner kicks. Evaluated by Google in collaboration with Liverpool Football Club, TacticAI's recommendations were found to be recognized by human expert evaluators at 90%. The outcomes have been published in a Nature subissue of Nature Communications. The paper is available at: https://www.nature.com/articles/s41467-024-4596...

VLOGGER generates photorealistic human videos containing facial and body movements via audio or text inputs, combined with individual images, using a stochastic diffusion model and a 3D human pose representation; a new large-scale diversity dataset, MENTOR, is introduced to provide 3D pose and expression annotations to support VLOGGER training, making it the largest dataset in terms of identity and temporal length; VLOGGER outperforms state-of-the-art methods on multiple public benchmarks, demonstrating advantages in image quality, identity retention, and temporal consistency, while validating its robustness across different diversity dimensions. (AI Mythic Room ...

Google DeepMind has developed an AI intelligence called SIMA, a generalized AI for 3D virtual environments.SIMA understands natural language commands and is able to perform tasks in different game worlds. Studies have shown that SIMA outperforms professional intelligences in nine different 3D games, demonstrating strong generalization across games. However, SIMA has not yet been able to reach the human level. (Heart of the Machine)

Several companies that provide technical support for cloud services and AI services are adjusting their expectations to their sales teams, emphasizing that the current hype around generative AI technology outweighs its actual capabilities, The Information exclusively reported. Executives, product managers, and salespeople at several major cloud providers, including Microsoft, Amazon, and Google, have also privately indicated that most of their customers are wary of investing in new AI technologies given the high price tag of AI services, fearing high costs, lack of accuracy, and difficulty in assessing the value of the technology. Some experts point out that while generative...

Google has released the MediaPipe LLM Inference API, which makes it easier for developers to run AI large-scale models locally on cell phones, PCs, and other devices. Google has focused on optimizing the cross-device stack, including new operations, quantization, caching, and weight sharing. Currently, MediaPipe supports four models, Gemma, Phi 2, Falcon and Stable LM, which run on web, Android and iOS devices. Google plans to extend this feature to more platforms. The demo ground...

On March 5, Google announced that it will be making some new changes to its search ranking system that will reduce search results for spammy, low-quality content. Pandu Nayak, Google's vice president of search, mentioned that Google is considering lowering its search rankings for low-quality articles that are created on a daily basis through low-paid contractors or AI generators.

Google is optimistic about the prospect of AI macromodels being used in smartphones. Brian Rakowski, vice president of product management for Google's Pixel division, recently predicted that Google's more advanced Gemini big models will be embedded in smartphones next year. (Punch)

Researchers at Google have proposed a new gated linear loop layer called the RG-LRU layer and designed a new loop block around it as an alternative to multi-query attention (MQA). They used this recurrent block to construct two new models: a model Hawk that combines MLP and the recurrent block, and a model Griffin that combines MLP, the recurrent block, and localized attention.By over-training Hawk and Griffin on 300B tokens for a range of different model sizes, it was found that Hawk-3B had a good performance on the downstream tasks of sexual ...

Google DeepMind introduces new architectures Hawk and Griffin, challenging traditional Transformer models and demonstrating the new potential of RNNs in AI.The Hawk and Griffin models demonstrate superior performance to Mamba at the same scale, proving their competitiveness in processing efficiency and downstream task performance. These two models achieve training efficiency comparable to Transformer and provide higher throughput and lower latency during inference, especially performing better when processing long sequence data.

Google CEO Pichai tackles racial controversy over Gemini image generation feature, promises structural reforms. Previously, Google suspended the feature due to problematic results from the tool, with Pichai stating that bias and offending users is an unacceptable mistake and that progress has been made on protective measures. (Tencent Technology)

At Mobile World Congress 2024, Google announced that Gemini will be integrated with Messages. This means that users will be able to access Gemini directly in Google Messages on their Android phones to chat, draft messages, and more with Gemini without leaving the app.

Google has released a new AI-based world model with 11 billion parameters. Using just one image, an interactive world can be generated, which is "motion-controlled" and in which users can act frame by frame. Google named the model Genie (Generative Interactive Environment, the word Genie means "Genie" in Chinese). Google said that Genie opens the era of "graphic/text generated interactive world", and will also become a catalyst for the realization of universal AI Agent.

Original link