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.
OpenAI Commits to Implementing Changes to Prevent Future Deception with ChatGPT
OpenAI has announced it will make changes to the way it updates the AI models that power ChatGPT following an incident where the platform became overly flattering towards many users.
Last weekend, after OpenAI launched the modified GPT-4o model – the default model powering ChatGPT – users on social media noticed that ChatGPT began responding in a way that excessively agreed with and validated their input. This soon turned into a meme as users shared screenshots of ChatGPT complimenting all kinds of questionable and dangerous decisions or ideas.
In a post on X (formerly Twitter), CEO Sam Altman acknowledged the problem and said OpenAI would work on finding solutions “as soon as possible.” On Tuesday, Altman announced the cancellation of the GPT-4o update and that OpenAI was working on “additional fixes” for the model’s behavior.
Following the incident, the company released a report on Tuesday and expanded on the specific changes it plans to implement regarding its model deployment process in a blog post on Friday.
Airbnb Launches AI-Powered Customer Service Chatbot in the US
Airbnb has begun rolling out its AI-powered customer service chatbot in the US last month, according to CEO Brian Chesky during a conference call for the company’s first-quarter earnings on Thursday.
Chesky said that 50% of Airbnb users in the US are already using the AI chatbot for customer service, adding that the company plans to roll out the feature to all US users this month.
Ai2’s New Small AI Model Claims to Outperform Similar Models from Google, Meta
The nonprofit AI research institute Ai2 released its Olmo 2 1B model on Thursday, claiming that it outperforms similar models from Google, Meta, and Alibaba in various benchmarks. The 1B model, which has a billion parameters, is now available under the permissive Apache 2.0 license on the AI development platform Hugging Face.
Unlike most models, the Olmo 2 1B can be reproduced from scratch, as Ai2 has provided the code and datasets used to train the model (Olmo-mix-1124 and Dolmino-mix-1124).
While smaller models like this might not match the performance of their larger counterparts, they don’t require heavy hardware to run, making them accessible to developers and hobbyists who are working with limited resources.
Anthropic Allows Users to Integrate More Apps with Claude
Anthropic launched on Thursday a new feature that allows users to connect more apps and tools to its AI chatbot Claude. Additionally, the company rolled out an extended “deep search” feature, enabling Claude to search the web, institutional accounts, and more.
These new features—called “integrations” for apps and “advanced search” for the deep search tool—are available in the beta version for subscribers of Claude’s Max, Team, and Project plans, with plans to include the Pro plan soon. Anthropic also raised the price limits for its AI-powered coding tool, Claude Code, for Max customers.
Microsoft’s New Phi 4 AI Model Competes with Larger Systems
Microsoft launched several new “open” AI models on Wednesday, with the most powerful being the Phi 4 series. The new Phi models, including Phi 4 Mini, Phi 4, and Phi 4 Plus, claim to compete with OpenAI’s O3-mini on at least one metric.
As described, all of these models are “inference” models, meaning they are designed to validate complex problem-solving processes over extended periods. They expand the Phi series, which Microsoft introduced last year to support AI developers creating edge applications.
The Phi 4 Mini, weighing in at 3.8 billion parameters, was trained on synthetic math problems and is intended for use in educational applications, such as integrated tutoring on lightweight devices.
Meta Predicts $1.4 tln in Revenue from Generative AI by 2035
Meta predicts that its generative AI products will generate between $2 billion and $3 billion in revenue in 2025 and between $460 billion and $1.4 trillion by 2035, according to court documents revealed earlier.
The documents, submitted by the attorneys representing authors suing Meta for allegedly using their works without permission to train AI models, don’t specify what Meta considers “generative AI products.” However, the company is known to be profiting—and expects to profit more—from generative AI in various forms.
Meta has agreements to share revenue with companies that host its open LLaMA models, and it recently launched an API for customizing and evaluating these models. Meta may also integrate AI features into its own advertising and subscription services, according to CEO Mark Zuckerberg during a Q1 earnings call.
Amazon Unveils Nova Premier, Its Most Advanced AI Model Yet
Amazon introduced on Wednesday what it claims to be the most efficient model in its Nova family, the Nova Premier.
Available on Amazon’s Bedrock AI platform, Nova Premier is capable of handling text, images, and video, but not audio. Amazon asserts that it excels in complex tasks that “require deep context understanding, multi-step planning, and precise execution across tools and data sources.”
While Nova Premier is weaker in some benchmarks compared to leading AI models from competitors like Google, it performs well in Amazon’s internal tests for knowledge retrieval and visual understanding.
Google’s Gemini Chatbot Now Allows Image Editing
Google announced on Wednesday that its Gemini chatbot can now edit AI-generated images as well as photos uploaded from users’ devices.
This new image editing feature is gradually being rolled out and will be available in most countries, supporting over 45 languages in the coming weeks.
Duolingo Launches 148 AI-Generated Language Courses
Duolingo announced the launch of 148 new AI-generated language courses. This release comes amid criticism following the company’s announcement that it plans to replace contractors with AI, positioning itself as an AI-first company.
According to Duolingo, this release doubles the current course offerings and marks the largest expansion of content in the company’s history. CEO Luis von Ahn emphasized that this launch showcases the tremendous impact of generative AI, which has allowed Duolingo to scale rapidly with unprecedented quality.
DeepSeek Updates Its Prover AI Model Focused on Mathematics
Chinese AI lab DeepSeek has quietly updated its Prover model, designed to solve mathematical proofs and theories.
The new version, Prover 2, has been uploaded to the Hugging Face platform and is based on the company’s V3 model, which contains 671 billion parameters and uses a mixture of experts (MoE) architecture.
Google Launches New AI Tools for Language Learning through Personalized Lessons
Google has launched three new AI-powered experiences to help users learn languages more effectively. While still in early testing, these features may position Google to compete with Duolingo.
The first experience helps users quickly learn phrases relevant to their immediate needs, the second focuses on helping users sound more like native speakers, and the third uses the camera to teach new words based on the user’s surroundings.
Meta Launches Standalone AI App to Compete with ChatGPT
After integrating Meta AI into WhatsApp, Instagram, Facebook, and Messenger, Meta has launched a standalone AI app. The app was revealed at Meta’s LlamaCon event and gives users access to Meta AI in a single application, similar to ChatGPT and other AI-powered assistants.
What differentiates Meta’s app is its ability to leverage information shared by users on Meta platforms like Facebook and Instagram to offer more personalized responses.
Alibaba Unveils Qwen3, a Family of Hybrid AI Models
Alibaba has launched Qwen3, a new family of hybrid AI models that it claims can match or even outperform the best models from Google and OpenAI.
The models, ranging in size from 0.6 billion to 235 billion parameters, are available for download or will soon be available under an open license on platforms like Hugging Face and GitHub.
Alibaba’s hybrid models take more time to reason through complex problems but perform quicker on simpler tasks.
Jensen Huang Confirms US and China Are “Very Close” in the AI Chip Race
Jensen Huang, CEO of Nvidia, confirmed that the U.S. and China are extremely close in the race for AI chip dominance, adding that America needs to implement energy policies to accelerate emerging tech industries.
Huang stated that China is not far behind the U.S. and, in fact, “is right behind us.” Nvidia, which produces some of the most in-demand AI chips, has faced restrictions from the U.S. government, which has imposed licensing requirements on sales to China.
Orb Mini Device Launched to Differentiate Between Humans and AI Online
In a new development for the global human verification project, Sam Altman’s startup “Tools for Humanity” unveiled a mobile device called Orb Mini, designed to differentiate between humans and AI online.
The device, which was revealed during the “Finally” event in San Francisco, features dual front-facing sensors for iris scanning to create a unique digital identity for the user.
Apple and Anthropic Collaborate to Develop AI-Powered Coding Platform
Apple has partnered with the AI startup Anthropic, backed by Amazon, to develop a new “vibe-coding” programming platform that uses AI to write, edit, and test code on behalf of programmers.
The new AI-powered system, which is an upgraded version of Apple’s Xcode, will integrate Anthropic’s Claude Sonnet model for coding, according to reports by Reuters and Bloomberg.
