The artificial intelligence (AI) realm witnesses rapid developments around the world, reshaping the world of digital transformation and accelerating the pace of AI adoption. In this weekly newsletter, FollowICT is highlighting the most important news and advancements in AI around the globe.
SoundCloud Changes its Policies to Allow AI Training on User Content
SoundCloud has quietly updated its terms of service to allow the company to train AI on the audio uploaded by users to its platform.
As noted by technology ethics expert Ed Newton-Rex, the latest version of SoundCloud’s terms includes a clause granting the platform permission to use uploaded content “to inform, train, or develop” artificial intelligence.
“You expressly agree that your content may be used to inform, train, develop, or provide inputs to AI or machine learning technologies or services as part of the services provided,” the terms state, which were last updated on 7 February, according to TechCrunch.
The terms include an exception for content under “separate agreements” with third-party rights holders, such as record labels. SoundCloud has a number of licensing agreements with independent production companies as well as major music publishers, including Universal Music and Warner Music Group.
Meta Appoints Ex Google DeepMind Director to Lead its AI Research Lab
Meta has appointed Robert Fergus to lead its Fundamental AI Research (FAIR) lab, according to Bloomberg.
Fergus served as a research director at Google DeepMind for nearly five years, according to his LinkedIn profile. Before joining Google, he was a scientific researcher at Meta.
Meta’s FAIR center, which has been operating since 2013, has faced challenges in recent years, according to a report from Fortune. FAIR led the company’s early AI model research, including Llama 1 and Llama 2. However, reports indicated that researchers left the unit in large numbers to join other startups, even to Meta’s newer GenAI group, which led the development of Llama 4.
Google Launches “Implicit Caching” to Make Access to the Latest AI Models Cheaper
Google is rolling out a feature in its Gemini API that the company claims will make its latest AI models cheaper for external developers.
Google calls this feature “implicit caching,” and it says it saves 75% of the “repeated context” passed to the models through the Gemini API. The feature supports Google’s Gemini 2.5 Pro and 2.5 Flash models.
Caching, a common practice in AI, reuses frequently used or precomputed data from the models to reduce computing demands and costs. For example, cached memory can store answers to questions that users frequently ask the model, saving the model from having to regenerate answers for the same request.
Google Launches AI Tools to Protect Chrome Users From Fraud
Google has announced new AI-powered defenses against fraud for its Chrome browser. The tech giant will begin using large language models (LLM) on desktops to protect users from online scams. It will also launch new AI-driven warnings for the Chrome browser on Android to help users stay alert to unwanted notifications.
According to Google, the enhanced Safe Browsing mode in Chrome provides the highest level of protection, offering double the protection against phishing and other online threats compared to the standard Safe Browsing mode. Google will now use Gemini Nano to provide users with enhanced protection against online fraud.
Amazon Designs New AI Tool to Improve Product Listings
Amazon has announced that it will launch a new AI tool to help sellers optimize their listings that contain missing details or attributes.
Sellers often list hundreds of products on Amazon and need to update information about them from time to time, which can be a cumbersome task. Therefore, the e-commerce giant wants sellers to use its new AI-powered tool, called “Optimize My Listing,” to make this process easier.
The tool automatically suggests product titles, attributes, descriptions, and missing details based on seasonal trends. Sellers can accept, reject, or modify these suggestions before updating their listings in Amazon’s catalog.
Fastino Trains AI models on Gheap gaming GPUs, Raises $17.5 MLN
Fastino, a startup based in Palo Alto, California, claims to have created a new type of small, task-specific AI model architecture. Fastino says these models are small enough to be trained on low-cost gaming GPUs, totaling no more than $100,000 in value.
This method is drawing increasing attention. Fastino has raised $17.5 million in seed funding led by Khosla Ventures, known for being an early investor in OpenAI.
With this funding, the total raised by the startup is nearly $25 million. The company had raised $7 million in November during a pre-seed funding round led by Microsoft’s M12 venture capital arm and Insight Partners.
Fastino has developed a range of small models it sells to its business customers. Each model focuses on a specific task a company might need, such as sensitive data editing or summarizing company documents.
Anthropic Launches AI-powered API for Web Search
Anthropic is launching a new API that enables its AI models, “Claude,” to search the internet. In a press release published Wednesday, the company said developers using this API can build applications powered by Claude that provide updated information.
The release of the API comes as AI companies look to enhance their models in various ways to attract new customers to their platforms. Last week, Anthropic also launched a tool to connect apps to the Claude platform, along with an expanded “Deep Search” feature that allows Claude to search company accounts, websites, and more.
Anthropic’s statement reads: “Developers can now enrich Claude’s extensive knowledge with current, real-world data by enabling web search while making requests to our API. With web search, developers can now build AI solutions that leverage up-to-date information without managing their own web search infrastructure.”
Microsoft Adopts Google’s Standard for Connecting AI Agents
Microsoft says it is adopting the open protocol recently launched by Google to allow AI “agents” to communicate with each other.
On Wednesday, Microsoft announced it would support Google’s Agent2Agent (A2A) specifications on two of its AI development platforms, Azure AI Foundry and Copilot Studio. Microsoft has also joined the A2A working group on GitHub to contribute to the protocol and tool development.
By supporting A2A and building on our open coordination platform, we are laying the groundwork for the next generation of software—collaborative, observable, and adaptable by design,” the company wrote in a blog post. “The best agents won’t reside in one app or cloud but will work across workflows, extending across models, domains, and ecosystems.”
The A2A protocol, unveiled by Google in early April, allows agents—semi-independent programs powered by AI—to work together across different clouds, apps, and services. Using this protocol, agents can exchange goals and execute actions. Developers are provided with a suite of compatible components they can use to ensure the agents collaborate securely.
Mistral Launches New AI Model Claiming Leading Performance for its Price
French AI startup Mistral has launched a new AI model, Mistral Medium 3, which focuses on efficiency without compromising performance.
Mistral Medium 3 is available via the Mistral API at a price of $0.40 per million input tokens and $2 per million output tokens. It delivers performance equal to or exceeding 90% of the more expensive Claude Sonnet 3.7 model from Anthropic across all performance benchmarks, according to Mistral. It also outperforms newer open models, including Meta’s Llama 4 Maverick and Cohere’s Command A, in popular AI performance evaluations.
Hugging Face launches Free AI Agent Tool Similar to Operator
A team from Hugging Face has released a free, cloud-hosted AI-powered tool. But beware, it’s very slow and occasionally makes mistakes.
Hugging Face’s agent, named Open Computer Agent, is accessible online and can use a pre-loaded virtual Linux machine with many applications, including Firefox. Similar to OpenAI’s Operator program, you can ask Open Computer Agent to perform a task; e.g., “use Google Maps to find the headquarters of Hugging Face in Paris”—and then relax as the agent opens the necessary apps and identifies the required steps.
Relevance AI Raises $24 MLN to Help Businesses Build AI Agents
Relevance AI, a startup based in San Francisco and Sydney, working on developing an “operating system” for AI agents, has raised $24 million in a Series B funding round led by Bessemer Venture Partners. Returning investors, including King River Capital, Insight Partners, and PeakX Ventures, also participated, bringing Relevance AI’s total funding to $37 million. The company has not disclosed its valuation.
This funding comes nearly a year and a half after the startup closed its Series A round. Relevance says it has experienced rapid growth, with 40,000 AI customers registered on its platform by January 2025.
Relevance says it will use the new funding to enhance its AI agent product capabilities and support customers in its key markets, Australia and the United States. The company has expanded to San Francisco and has a growing team to support its market rollouts. Relevance now has 80 employees across its San Francisco and Sydney offices, up from 19 employees in 2023.
Alongside its Series B funding, Relevance AI is launching two new features on its platform. “Workforce” is a no-code multi-agent system designed to help non-technical professionals and engineers build specialized teams of agents to collaborate like human employees and complete complex tasks from start to finish. “Invent” is a tool that enables users to create AI agents using text prompts.
Alibaba Innovates Method to Reduce AI Training Costs for Search by 88%
Researchers at Alibaba Group in China have developed a new approach that can reduce the cost and complexity of training AI systems for information retrieval, eliminating the need for expensive commercial search engine APIs.
This technology, named “ZeroSearch,” allows large language models to develop advanced search capabilities through simulation instead of interacting with real search engines during the training process.
This innovation could save companies significant expenses previously spent on APIs while providing better control over how AI systems learn to retrieve information.
OpenAI Plans to Expand its Massive “Stargate” Project to Outside the US
The Financial Times reports that OpenAI plans to expand its massive $500 billion AI infrastructure project “Stargate” to include projects outside the United States as part of OpenAI’s initiative for foreign countries.
The paper adds that the company has begun marketing the initiative globally to some of the U.S.’s traditional allies, such as France, the UK, and Germany, who have shown initial interest in participating.
The Stargate project was officially launched in January at the White House with President Donald Trump, SoftBank founder Masayoshi Son, and Oracle CEO Larry Ellison in attendance. It currently centers around a data center under construction in Abilene, Texas.
New General AI Agent “Suna” competing with China’s “Manus”
Kortix AI recently launched “Suna,” which it describes as the first open-source general-purpose AI agent. The Texas-based software development startup said that the Suna agent is versatile and capable of performing independent actions and completing complex real-world tasks based on user input.
Suna competes with the Chinese general AI agent “Manus,” which caused a stir when it launched earlier this year due to its impressive capabilities, according to several reports.
Suna is designed for developers and individuals looking to boost productivity and automate certain tasks. The company describes it as a “smart employee powered by AI” that can carry out multi-step tasks on behalf of the user.
Chinese Baidu Seeks AI Patent for Decoding Animal Sounds
Ever wished you could understand what your cat is trying to say? Chinese tech company Baidu is exploring the possibility of using AI to translate its mysterious meows into human language.
Baidu, the owner of China’s largest search engine, has filed a patent with the National Intellectual Property Administration in China for a system that translates animal sounds into human language, according to a patent document published this week.
Scientists have long attempted to decode animal communication, and Baidu’s patent represents the latest effort to harness AI for this task.
The document states that the system will collect animal data, including vocal sounds, behavioral patterns, and physiological signals, which will be pre-processed and integrated before being analyzed using AI designed to recognize the emotional state of the animal.
The emotional states will then be linked to semantic meanings and translated into human language.
Amazon Works on AI Tool “Kiro” using agents to simplify programming
According to an internal document obtained by Business Insider, Amazon Web Services is building a new AI programming tool called Kiro.
The software development tool leverages AI agents to analyze user prompts and existing data to generate code “almost in real-time,” the document states.
Kiro is a web and desktop app that can be customized to work with third-party AI agents, according to Amazon’s description. It also utilizes knowledge bases, plugins, and topics to enhance developer productivity.
Kiro will feature a multimedia interface, allowing developers to input not only text but also visual diagrams and other contextual information, according to the document.
