Hugging Face has recently introduced Transformers Agent, which allows customers to control over 100,000 HF models using the Transformers and Diffusers interface. Other HF models can be generated using the new API of tools and agents to meet complicated, multimodal issues.
Transformers Agent is a natural language API built on top of transformers that includes a set of curated tools and an agent meant to read natural language and use these tools. The system is designed to be extendable, with the flexibility to simply add any relevant community tools.
Users must first create an agent, which is an LLM, before they can utilize the agent run functionality. Both OpenAI modes and open source alternatives like BigCode and OpenAssistant are supported by the system. Hugging Face provides free access to BigCode and OpenAssistant APIs.
The tools are made up of a single function with a specific name and description that is then used to prompt an agent to undertake a specific activity. The agent is taught how to use these tools with a prompt that shows how the tools can be used to complete the desired inquiry. While pipelines usually integrate numerous jobs into a single operation, tools are designed to focus on a single, straightforward action.
There are two APIs available, Single execution (run) and Chat-based Execution (chat). The agent’s run() method is available to the user as a single execution method. This method automatically identifies and executes the relevant tool or tools for the task at hand. The chat() method is used by the agent in its chat-based approach. This approach is especially beneficial when the state must be maintained over multiple instructions.
Founders of Anthropic, who are also former employees of OpenAI, are attempting to make artificial intelligence safe. In an interview, Anthropic co-founder Jared Kaplan concentrated on an approach known as “constitutional AI” to achieve this.
He claims that this approach aims to train chatbots and other AI systems to adhere to predetermined laws or constitutions.
Historically, the development of chatbots like ChatGPT has relied on human moderators to assess the system’s performance for toxicity and hate speech. The system then makes adjustments to its answers using this feedback. Reinforcement Learning through Human Feedback, or RLHF, is the term used to describe this process. However, with constitutional AI, the chatbot itself is mostly in charge of this work. However, an individual is still needed for further analysis.
Anthropic has long discussed constitutional AI and trained its own chatbot, Claude, using the technique. The corporation is now disclosing the exact written rules it uses for such operations, which are found in its constitution. A number of sources, including the Universal Declaration of Human Rights and Apple’s terms of service, are cited in this paper. The goal of many of them is to avoid being rude.
While there are many unanswered questions in this situation, Kaplan emphasizes that his company’s goal is to demonstrate the overall effectiveness of its approach which is the notion that constitutional AI is superior to RLHF when it comes to managing the systems’ input.
An online machine learning course for students to learn about bandit algorithms has been announced by IIT Bombay. On the SWAYAM NPTEL platform, the course is called Bandit Algorithm (Online Machine Learning). Participants who wish to receive a certificate must pay Rs. 1,000 and pass an in-person examination at NPTEL centers, despite the course’s free enrollment and learning opportunities.
The course is being led by IIT Bombay teacher Manjesh K Hanawal, who has a MS degree in gadgets and correspondence designing from the Indian Establishment of Science, Bangalore and PhD degree from INRIA, Sophia Antipolis, France, and the College of Avignon, France. Additionally, he has been honored with SERB’s Early Career Research Award and DST’s Inspire Faculty Award.
Bandit algorithms that strike a good balance between exploration and exploitation in a variety of random environments will be taught to course enrollees. Some of the topics that will be covered are:
Introduction to stochastic setting and various regret notions.
Algorithms for pure explorations.
Introduction to bandit algorithms.
Regret lower bounds for adversarial setting.
Stochastic bandit algorithms.
Adversarial setting with full information and bandit information.
The course will be conducted from 24 July to 13 October 2023. The last date to enroll is 31 July 2023. The course also carries 3 credit points, which students can submit in their respective colleges. The examination to obtain the course certificate will be taken on 28 October 2023.
While the course is open to anyone interested, it is most suitable for postgraduate students from streams such as computer science, electrical engineering, operations research, mathematics, and statistics. Participants must have knowledge of the basics of probability theory and optimization.
Google just wrapped up its Google I/O keynote day, during which it unveiled some very significant changes to its Pixel portfolio and demonstrated the most recent developments in AI. Here is a list of all of the most significant announcements that was announced during the Google I/O 2023.
Search Generative Experience
A significant change to Google Search is coming in the shape of something called AI snapshots.
Users will begin to see AI-powered answers at the very top of their search results for select queries after they choose to use the new Search Generative Experience (SGE) feature, which can give the search additional context. The new PaLM 2 large language model (LLM), which Google also unveiled at Google I/O 2023, is the engine behind these AI snapshots. The new model, which has advancements in reasoning, coding, and translation, is already powering 25 Google services, including the company’s Bard chatbot.
Google Bard Public Access
At Google I/O 2023, Google declared that anyone can now use its AI-powered chatbot, regardless of whether they were on the queue. A number of additional features are also being added by the company, including support for the Japanese and Korean languages and a simpler method for exporting created text to Google Docs and Gmail.
In the future, Google aims to offer more capabilities in addition to the current dark mode and visual search tools. That includes interfaces with third-party businesses like OpenTable and Instacart as well as AI image generation using Adobe’s AI image generator, Firefly.
Google has said that it will also be delivering new Android features that are powered by AI. One of these capabilities, called Magic Compose, will be available in the Messages app for Android and allow users to reply to texts using AI-generated suggestions for words.
Additionally, a feature that enables users to design their own wallpaper using AI is being added. Soon, users will be able to describe an image, and the device will create it using Google’s text-to-image diffusion technique, as opposed to having to select from a list of predefined possibilities.
Duet AI
Google also revealed Duet AI, a new moniker for the AI toolkit that it is introducing to Docs, Sheets, Slides, Meet, and Gmail during the Google I/O 2023. These Workspace capabilities, according to the announcements, allow users to do things like create graphics from text in Slides or compose emails, but they are presently only accessible to people who sign up for its waitlist.
Google Photos Magic Editor
New AI-powered editing tools are coming to Google Photos that will allow one to make some pretty big changes to a photo, such as brightening the sky, repositioning a person or item, and removing background subjects. Later this year, a limited number of Pixel customers will receive early access to this experience called Magic Editor.
Codey
Codey, a brand-new tool for code completion and generation, was also announced at Google I/O 2023. It’s essential for various man-made intelligence driven coding instruments being sent off today and is likewise Google’s solution to GitHub’s Copilot, a talk device utilized for posing inquiries about coding. As per the announcements, Codey has been trained to respond to prompts related to coding as well as general Google Cloud-related questions.
Pixel Fold, Pixel Tablet, and Pixel 7A
With the introduction of the Pixel Fold, Pixel Tablet, and the reasonably priced Pixel 7A, Google’s Pixel lineup took center stage throughout the Google I/O 2023. Google’s first foldable device, the $1,799 Pixel Fold, has a 5.8-inch exterior OLED screen that unfolds to show a bigger 7.6-inch display that is also an OLED panel with a 120Hz refresh rate. Also available is the $499 new 11-inch Pixel Tablet. The device’s magnetic charging dock, which also serves as a speaker, allows it to work as both a standard tablet and a smart display.
Moreover, Google added the Pixel 7A to its midrange A-series lineup. The device includes Google’s Tensor G2 chip, along with a 6.1-inch 1080p display that can operate at up to 90Hz, just like the Pixel Fold and Pixel Tablet.
Wear OS 4
According to the announcements, Google is preparing Wear OS 4 even though Wear OS 3 hasn’t yet finished rolling out to all users. The revised OS will have better battery life, a mechanism for customers to back up and restore their watches, and new accessibility capabilities. Smart home integrations are also added to Wear OS 4, enabling you to manage media, adjust lighting, and view animated camera notification previews all from your watch. For the time being, Wear OS 4 is only accessible through a developer preview and emulator. Full availability is expected later this year.
Google Home App
On May 11, the revamped Google Home app will stop being a public preview that is only accessible by invitation and will finally be open to all users. A new Favorites page, an improved camera interface, and compatibility for a wide range of new device types are just a few of the significant upgrades to the redesigned app.
A digital engineering company called Objectivity, which specializes in cloud and platform development services and aids clients in accelerating their transformation journeys for quick innovation, has agreed to be acquired by Accenture. The deal’s financial details were not made public.
The highly qualified team from Objectivity, with its headquarters in Coventry, UK, will join Accenture, providing size and resources from Objectivity’s operations in the UK, Poland, Germany, and Mauritius.
Accenture’s Cloud First capabilities will be strengthened by Objectivity’s strong platform engineering, cloud native computing, and application modernization expertise. This will enable clients to take advantage of more flexible platform models and services to hasten the digital transformation of their businesses.
“The world is undergoing enormous change, necessitating the constant reinvention of organizations in order to create value. The depth and breadth of Objectivity’s technical resources and expertise will strengthen our European platform engineering capabilities, giving our clients the agility they need to reimagine their businesses and industries,” said Jean-Marc Ollagnier, CEO of Accenture for Europe.
Rob Helle, the company’s creator, commented, “For more than 30 years, Objectivity has supplied original thinking to develop the most purpose-fitting digital solutions. Our values-driven culture has effectively helped us grow into an international, virtual business where clients and staff have been expanding alongside one another. We will be able to accomplish more since we will be stronger as a team with Accenture.”
Startup UVeye, which develops automated vehicle inspection systems, announced today that it has closed a $100 million fundraising round headed by Hanaco VC.
The company has now collected about $200 million in total, thanks to the Series D round, which also included current investors W.R. Berkley Corporation, GM Ventures, CarMax, F.I.T. Ventures L.P., and numerous Israeli institutional investors.
A drive-through system developed by UVeye uses artificial intelligence and sensor fusion technologies to quickly identify mechanical and exterior defects below or on the sides of any car. Additionally, it may spot changes and foreign factors that might be problematic for the vehicle.
According to the firm, as electronic and driverless vehicles get more complex, a system like this is required. The capacity to undertake low-cost and high-frequency predictive maintenance will become crucial for businesses that run fleets of these cars, it claims.
The co-founder and CEO of UVeye, Amir Hever, stated that the company’s objective is to standardize how the car industry identifies damage and mechanical problems on automobiles and to develop new quality standards. He said, “Our patent-protected technology delivers unmatched solutions for swiftly and precisely identifying vehicle problems to automakers, dealers, and fleet operators.”
According to sources, Spotify, the biggest player in the audio streaming industry, recently removed tens of thousands of songs, or roughly 7% of the recordings that the platform Boomy had uploaded.
According to an individual familiar with the incident, the major record label Universal Music had alerted all of the major streaming services of unusual streaming activity on Boomy recordings.
The Boomy songs were taken down due to suspicions of “artificial streaming,” which is online bots impersonating real people to increase the number of listeners for particular songs. Because technology enables the instant generation of several music files that can then be posted online and streamed, AI has made this kind of activity simpler.
Boomy, a platform that was introduced two years ago, allows users to select different genres or descriptors, such “rap beats” or “rainy nights,” to produce music that is automatically generated. The song can then be made available via streaming platforms, where users will get royalties. According to Boomy, a company based in California, its customers have written over 14 million songs.
According to the firm, “artificial streaming is a long-standing, industry-wide problem that Spotify is working to stamp out across our service.”
Last month, according to a report in The Financial Times, Universal Music wrote to streaming services requesting that they take action to limit the use of generative AI on their platforms. The Weeknd and Drake’s voices were imitated in a song that went viral on streaming services the same week.
ImageBind, an open-source AI model that can simultaneously learn from six different modalities, has been released by Meta. Machines can now comprehend and link various types of data, including text, image, audio, depth, temperature, and motion sensors. Without having to be taught on every potential modality combination, machines can learn a single shared representation space using ImageBind.
ImageBind is significant because it gives machines the ability to learn holistically. Researchers might investigate novel possibilities by fusing various modalities, such as developing multimodal search tools and building immersive virtual environments. By effortlessly generating richer media, ImageBind could help enhance content recognition and moderation while fostering creative design.
Meta‘s greater objective of developing multimodal AI systems that can learn from all kinds of data is reflected in the creation of ImageBind. Researchers now have additional options to create fresh, all-encompassing AI systems, thanks to ImageBind, as the number of modalities rises.
AI models that rely on many modalities have a lot of room to grow because of ImageBind. ImageBind learns a single joint embedding space from image-paired data that enables several modalities to “talk” to one another and discover relationships without being observed simultaneously. This makes it possible for other models to comprehend novel modalities without the need for time-consuming training.
A larger vision model may be advantageous for non-visual tasks like audio classification because of the model’s strong scaling behavior, which shows that its performance increases with the strength and size of the vision model. Along with audio and depth classification tasks, ImageBind performs better than earlier research in zero-shot retrieval.
At its annual Think conference today, IBM unveiled IBM Watsonx, a new AI and data platform that will allow businesses to scale and hasten the impact of AI with reliable data.
Enterprises utilizing AI today require access to a complete technology stack that enables them to develop, test, and roll out AI models across their organization with reliable data, speed, and governance, all this in one location and across any cloud environment.
With Watsonx, IBM is providing businesses with a seamless end-to-end AI workflow that will make AI easier to adapt, scale, and access foundation models that have been curated and trained by IBM as well as open-source models. It will act as a data store to enable the gathering and cleaning of training and tuning data, and a toolkit for governance of AI.
Arvind Krishna, chairman and CEO of IBM, said, “With the development of foundation models, AI for business is more powerful than ever. Using AI is vastly more scalable, economical, and effective when foundation models are used.”
“In order for clients to be more than just users and gain the benefits of AI, we designed IBM Watsonx specifically for the needs of businesses. While maintaining complete control over their data, companies can swiftly train and deploy customized AI capabilities across their whole business with IBM Watsonx,” he added.
Clients will have access to the set of tools, technology, infrastructure, and consulting skills needed to develop their own AI models on their own data, or to enhance and modify existing models, and then deploy them at scale in a more dependable and open environment to promote commercial success.
Chinese police detained an individual on Sunday in what may be the first instance of someone reportedly spreading false information using the hugely popular AI chatbot ChatGPT.
According to the South China Morning Post, authorities in the northern Chinese province of Gansu say they detained a man for allegedly fabricating news reports about a train accident that claimed nine lives.
The man, whose last name was given as Hong by the authorities, was said to have created the “fake” news using ChatGPT and propagated it via numerous blogs online. Hong was expressly accused of “picking fights and causing trouble,” a general political charge leveled at activists and dissidents.
Several hundred users on the Chinese blogging platform Baijiahao are said to have posted the disputed posts. Police allegedly claimed that Hong circumvented Baijiahao’s rules on submitting the same content to several accounts. He used ChatGPT to create various versions of the news article, resulting in about 15,000 clicks.
Hong apparently owns a business that runs numerous blog-style websites. According to reports, the platforms are registered in Shenzhen as a significant big tech manufacturing and business center in southern China.
