Apple has announced the opening of Siri AI to all developers, starting with the betas of iOS 27.3, iPadOS 27.3, macOS 27.3, and visionOS 27.3. This initiative allows developers to integrate their app's content and actions directly into Siri AI using App Intents. Siri AI is described as an entirely new, profoundly more capable, conversational, and personal version of Siri, designed for deep integration across Apple's operating systems and everyday apps.

This opening enables users to interact with apps through Siri AI using natural language for custom actions, such as checking deliveries, booking reservations, or placing orders. For complex requests, new "skills" are introduced, allowing Siri AI to combine several actions into a single experience and surface any entity within the Spotlight semantic index.

Apple has already rolled out these capabilities across many of its own apps, including Calendar, Mail, and Maps, providing developers with examples to reference. This early Developer Preview offers developers a chance to explore, build, and shape their experiences before the public launch in a few months and App Store submissions open in the spring.

👉 Read the full article on Apple Developer

A team of researchers at MIT successfully demonstrated that GPT-6 Astra, an advanced OpenAI model, could identify and correct an internal flaw in another AI, Alibaba's Qwen3.5-0.8B. Remarkably, this improvement was achieved without retraining Qwen, by modifying only 0.016% (131,072 weights) of its approximately 800 million parameters.

The experiment identified that Qwen was getting stuck in reasoning loops while solving mathematical problems. Astra pinpointed the components associated with this behavior and directly modified the parameters using a method called Weight-Space Programming. This intervention dramatically increased Qwen's accuracy on the MATH-500 test from 7.9% to 48.1%, on GSM8K from 13.8% to 67.1%, and on IFEval from 21.0% to 62.3%.

This research highlights a significant potential for AI to correct specific flaws in other AI models without extensive retraining, paving the way for more efficient model development and targeted interventions for problematic behaviors. However, it also raises concerns about recursive self-improvement scenarios where AIs contribute to creating progressively more capable systems, and the human ability to oversee the scope of changes AIs make to other models.

👉 Read the full article on Gizmodo en Español

ChatGPT is claiming to handle "search-like queries" equivalent to about 20% of Google's search volume, effectively positioning itself as the world's second-largest search engine, despite Google's over 90% market share. Adam Fry, OpenAI's ChatGPT consumer product lead, defines "search-like queries" as any question a user would traditionally go to a search engine for but now chooses ChatGPT.

While Fry didn't provide specific numbers, based on Google's reported 16 billion searches per day, ChatGPT is estimated to be handling around 3.2 billion similar user interactions daily. Furthermore, ChatGPT reportedly boasts 1.2 billion weekly active users, with an addition of 200 million over the last two months.

The platform's models are also generating approximately 1.5 billion images per week, and about 250 million people are utilizing ChatGPT Voice. However, Fry did not provide a concrete update on the number of paying subscribers, only stating it has grown from the last public figure of 50 million, highlighting the importance of scrutinizing what metrics companies choose to share and withhold.

👉 Read the full article on Gizmodo