Leap towards action
Large language models (LLMs) are all the rage at the moment for their ability to understand and generate text. But they are not quite adept in doing tasks that may require real-world simulations. Here comes large Action Model (LAM), a substantial leap in the world of artificial intelligence (AI) that can facilitate AI systems to do complex tasks based on human instructions. LAM has been recently created by Microsoft researchers that can operate Windows progammes on its own. This represents a huge shift from the existing AI models – from conversation to performing tasks. As compared to existing AI models, LAM can understand prompts and commands in text, voice and images and turn them into detailed real actions and can also adjust approach in real time which makes them capable of not just understanding but acting upon it as well. LAMs work on two kinds of data — task-plan data which are essentially about tasks such as opening a document and highlighting text and task-action data which are specific steps to do a certain task.
AI agents in workforce
In a recent blog post titled Reflections, OpenAI CEO Sam Altman said the company knows how to build artificial general intelligence (AGI) “as we have traditionally understood it”. “We believe that, in 2025, we may see the first AI agents join the workforce and materially change the output of companies,” he added. Altman is of the view that iteratively putting great tools in the hands of people leads to great, broadly-distributed outcomes. Declaring that OpenAI cannot be a “normal company”, Altman also talked about “superintelligence” and said it can massively accelerate scientific discovery and innovation in the next few years.
5-minute news
Google is set to launch a new AI feature called ‘Daily Listen’, which gives users a five-minute audio overview of important news stories. The function creates a customised audio overview based on a user’s discover feed and news searches, which includes a transcript and cover art.
Currently, users in the United States can access ‘Daily Listen’ by enabling it in the Google app’s search labs section. This breakthrough comes after Google launched NotebookLM’s Audio Overviews in September last year, which transforms documents into 10-minute podcasts with virtual hosts. Google has recently also given an upgrade to the Audio Overviews feature in Notebook LM which enables users to join the conversation among virtual hosts by simply calling in the podcast. At the same time, they can also ask the host to elaborate something in a different manner and seek more information. “It’s like having a personal tutor or guide who listens attentively, and then responds directly, drawing from the knowledge in your sources,” Google wrote on its support page.
Error in Apple Intelligence
After the BBC complained to Apple that its AI feature has rewritten one of the headline of its report to inaccurately say that one shooting suspect Luigi Mangione had died by suicide, the company is tweaking its AI-powered summary tool. BBC has been trying to get Apple fix the problem as several AI-generated summaries of notifications containing BBC News reports were inaccurate. Apple Intelligence, the tech giant’s AI system, falsely claimed that Tennis legend Rafael Nadal had come out as gay. It had also falsely summarised BBC News app notifications on its story about the PDC World Darts Championship semifinal, claiming British darts player Luke Littler had won the championship a day before the actual tournament’s final in which Littler did go on to win. Apple acknowledged the flaws and has offered a software update that clarifies when text is generated by AI. Users can disable or adjust summaries in device settings. Apple Intelligence, which was released in beta in the UK in December, is now accessible on the iPhone 15 Pro, iPhone 16 series, and smartphones equipped with M1 CPUs or later. These blunders underscore the difficulties in ensuring correctness in AI-generated material.
Deepfake detector
The global cybersecurity company McAfee has recently launched its deepfake detector in an effort to address the growing incidence of AI-generated deepfakes and scams. The application, which is available on some Copilot+ PCs in English language, notifies users within seconds if AI-altered audio is detected in videos, protecting their privacy by processing data immediately on the device. According to a McAfee survey, 66% of Indians have been victims of deepfake frauds, with many falling victim multiple times. The product, priced at `499 standalone or `2,398 with McAfee+, uses huge datasets to detect fake material and ensure safe digital navigation. The company claimed that it does not collect or record users’ audio in any way and they can turn audio detection on or off whenever they need, making the user always in control. Talking about the product, Pratim Mukherjee, senior director of engineering at McAfee, said, “While not all AI-content is created with malicious intent, the ability to know if a video is real or fake helps consumers make smart and well-informed decisions.”
Meanwhile, the British government is mulling over creating and sharing sexually explicit “deepfakes” a criminal offence in Britain. Under the new offence to be introduced by the government, perpetrators could be charged and face prosecution for both creating and sharing these images and offenders would face fines and even jail time. “There is no excuse for creating a sexually explicit deepfake of someone without their consent,” the justice ministry said in a statement.
Detecting cancer
A new study by researchers from the University of Lubeck in Germany has found that AI can improve cancer detection rates by more than 17%. AI-backed radiologists obtained a cancer detection rate of 6.7 per 1,000 screenings, which is a 17.6% improvement against traditional readings.