Artificial intelligence (AI) has gained immense popularity in the modern technological era and is embraced by individuals around the globe. People are increasingly incorporating AI into their daily lives, prompting AI companies to frequently update their products.
Here is a list of top AI updates from last week:
OpenAI launches ChatGPT memory update for its pro users
Recently, OpenAI has updated its ChatGPT to improve its “memory” feature, which will help the chatbot remember past conversations to give more personalized and relevant answers.
For example, if you tell it you like certain foods or have specific preferences, it can use that information later to tailor its responses.
Right now, this upgraded feature is only available to ChatGPT Pro subscribers, who pay $200 a month. OpenAI’s CEO, Sam Altman, said that the firm will soon come to Plus subscribers ($20 a month) too, but there is no word yet on when or if free users will get it or not.
One can turn off the memory feature or use a “temporary chat” mode if you don’t want your conversations to be saved or remembered.
Google announces its new tools and updates at Google Cloud Next 2025.
At Google Cloud Next 2025, Google has announced new tools and upgrades to make its cloud services faster, smarter, and safer.
AI Improvements
Google has introduced better AI models, like Gemini 2.5, which can think faster and handle tasks more efficiently as compared to other AIs. They also launched tools to help businesses create AI agents for things like customer service or data analysis.
Infrastructure Upgrades
They disclosed a new chip called Ironwood TPU, which is super powerful for running AI tasks. They also expanded their global network (Cloud WAN) to make data transfer quicker and more reliable for companies.
Security Enhancements
Google has also introduced Google Unified Security, a system that uses AI to spot and stop cyber threats across networks, apps, and devices. It’s an intelligent system that safeguards businesses against cyber threats.
These updates are aimed at helping businesses to work smarter, save their time, and stay secure while using Google’s cloud technology.
Google launches Agent2Agent, to help AI programs in companies work tiogether.
Google Cloud launched Agent2Agent (A2A), a free-to-use system that helps different AI programs “talk” to each other, even if they’re made by different companies.
Agent2Agent (A2A) is built using common tech standards like HTTP and JSON-RPC, which makes it easy to connect with existing business IT systems. It has strong security features to keep data safe. A2A lets AI programs handle long tasks and share updates as they go, like giving real-time progress reports. Its main features include managing tasks, figuring out what each AI can do, and helping AIs work together.
Each AI has a “profile” called an Agent Card, written in a simple format (JSON), that lists its skills so other AIs can quickly find the right one for a specific job.
Meta launches its Llama 4 open-weight AI model.
The US tech giant Meta has launched its latest open-weight large language model (LLM), Llama 4. The firm claims that its newly launched model Llama 4, is an open-weight answer to its rival firms like DeepSeek, Google, OpenAI, etc.
Llama 4 is a new set of AI models that will make Meta AI smarter on apps like WhatsApp, Instagram, and Messenger. Two models, Llama 4 Scout and Llama 4 Maverick, can be downloaded from Meta’s Llama website or Hugging Face. They also showed off Llama 4 Behemoth, a special model that helps train the other Llama 4 models.
The publicly available models trained with weights are known as open-weight AI models. Further, the users can also run them locally without depending on cloud APIs. While available, the models may still have licensing limitations on modifications or commercial use.
DeepSeek is now working on building a self-improving AI
DeepSeek has teamed up with Tsinghua University of Belgium to upgrade its new AI model, called DeepSeek-GRM, by converting it into an all-knowing mean machine that requires no human assistance.
The firm is implementing a technique called reinforcement learning, which has worked well for training AI models to do specific tasks quickly.
But this method is tricky to use, especially for bigger, more general tasks.
To tackle this, DeepSeek has come up with something called “self-principled critique tuning,” or “self-improving AI.” The technique helps the AI to improve itself by figuring out what is right and what is wrong.
Samsung and Google collab to include Google’s AI technology in its new home AI robot.
Samsung Electronics and Google Cloud have collaborated to bring Google Cloud’s generative AI technology to Samsung’s new AI robot named Ballie.
Available to consumers this summer, Ballie is designed to be a home robot that will be able to engage in natural, conversational interactions to help users manage home environments, including adjusting lighting, greeting people at the door, personalizing schedules, setting reminders, and more.
For example, if you ask Ballie, “How do I look?” It can look at your outfit and suggest things like adding a bright shirt or trying a new accessory.
Further, Ballie can also help with your health and home. If you say, “I feel unwell today,” it uses Gemini’s connection to Google Search to find helpful tips from trusted sources.
It might offer ways to get more exercise, improve your sleep space, or track your sleep habits, making it a handy companion for daily life.
Nvidia has launched its new AI model, Llama 3.1.
Nvidia has created its new AI model, Llama-3.1-Nemotron-Ultra-253B-v1, which is free for anyone to use. It’s smaller than some competitors, with 253 billion settings (parameters), but it performs impressively, better than a rival model called DeepSeek R1 in many tests.
This model is excellent at reasoning, following instructions, and helping with tasks like answering questions or assisting users. They discussed this topic at their major tech conference in March. The firm has improved its new AI model by giving it extra training.
They also simplified the model’s knowledge using 65 billion pieces of data and kept training it with another 88 billion pieces to make it smarter. The training data came from places like FineWeb, Buzz-V1.2, and Dolma, which are big collections of information.
For the final touches, the firm has also used both public data and AI-generated examples to teach the model how to switch between different ways of thinking and responding and created this excellent AI model.
Also Read: Pakistan’s AI Revolution: Saving Mothers with Technology
