Today's AI news: Apple pushes Siri ambitions, brain-wave privacy fears, robot learning environments, and pharma's AI scepticism.
Apple CEO Pushes for a More Ambitious AI Strategy — Including a ChatGPT-Style Siri
Apple's chief executive is pressing the company's software leadership to pursue a significantly more ambitious AI roadmap, according to a Bloomberg report circulating today. The push reportedly includes exploring a transformation of Siri into a conversational agent comparable to ChatGPT, potentially as part of an iOS 28 shake-up. The move signals that Apple — long seen as cautious in its generative AI rollout — is feeling competitive pressure to close the gap with rivals who have already deployed capable AI assistants at scale. The story is generating substantial discussion, reflecting broad interest in whether Apple can mount a credible challenge in the AI assistant space.
Brain-Wave Headsets Are Harvesting Neural Data to Train AI — Without Telling Users
A Science magazine investigation, one of today's most-read stories, reveals that makers of consumer brain-wave headsets are collecting neural data from users and feeding it into AI training pipelines — often without clear disclosure. The report raises serious privacy concerns about a category of wearable devices that sits closer to human cognition than almost any other consumer technology. Unlike browsing history or location data, brain-wave signals can potentially reveal emotional states, attention patterns, and other deeply personal information. Researchers and privacy advocates quoted in the piece are calling for regulatory frameworks to catch up with the technology before the practice becomes entrenched.
Pharma CEOs Push Back on Silicon Valley's AI Drug-Discovery Hype
A Wall Street Journal report finds that chief executives at major pharmaceutical companies are sceptical of the sweeping claims made by Silicon Valley about AI's ability to cure disease and revolutionise drug discovery. While technology companies have heavily promoted AI as a transformative force in medicine, drugmaker leaders point to the long, complex and heavily regulated pathway from a computational prediction to an approved therapy. The piece is a useful counterweight to the relentless optimism that has characterised much AI-in-healthcare coverage, and suggests a gap remains between what AI researchers promise and what the industry believes is practically deliverable in the near term.
AllenAI Releases ToyBlocks: A Controlled World for Robot Learning
The Allen Institute for AI (AllenAI) has published ToyBlocks on Hugging Face — a controlled simulated environment designed specifically for robot learning and evaluation. Structured benchmark environments like ToyBlocks are increasingly important as robotics researchers seek reproducible, standardised conditions in which to train and compare AI agents. By providing a known, controllable world, the project aims to make it easier to isolate what a robotic learning system has actually learned versus what it has merely memorised from training conditions. Full technical details are available via the Hugging Face blog.
Cult.fit Uses OpenAI Codex to Build Internal Tools Across the Business
Indian fitness and wellness platform Cult.fit has published a case study with OpenAI describing how it is using Codex to help teams across multiple business functions build software tools and move faster. The example is notable because it illustrates Codex being deployed not just by engineering teams but across operational functions — consistent with OpenAI's broader pitch that AI coding assistants can democratise software creation within organisations. The case study offers a concrete, real-world data point for enterprises evaluating whether AI-assisted development delivers meaningful productivity gains outside purely technical roles.
AI Interpretability Research: A New Tool for Reading Language Model Internals
A paper published on arXiv introduces Maximally-Activating Example Metamodels (MAEMs) — a technique for interpreting directions inside a language model's residual stream by training a secondary model to generate text that strongly activates those directions. The researchers report that a Qwen 3.6-27B MAEM outperforms corpus search and a natural-language autoencoder on a range of held-out tasks, activates 89 per cent of held-out sparse autoencoder features, and can recover the payloads of LoRA backdoors. Crucially, because the generated exemplars can be verified with a standard forward pass on the subject model, the approach offers a grounded — rather than purely qualitative — method for scaling interpretability research. This kind of mechanistic transparency work is becoming increasingly important as AI systems are deployed in high-stakes settings.
Frequently Asked Questions
What is Apple planning to do with Siri in its new AI strategy?
According to a Bloomberg report, Apple's CEO is pressing the company's software leadership to pursue a more ambitious AI strategy, including exploring a transformation of Siri into a conversational agent comparable to ChatGPT, potentially as part of an iOS 28 overhaul.
Why are consumer brain-wave headsets raising AI privacy concerns?
A Science magazine investigation found that makers of consumer brain-wave headsets are collecting users' neural data and using it to train AI systems, often without clear disclosure. Because brain-wave signals can reveal emotional states and attention patterns, researchers and privacy advocates are calling for regulatory action.
What do pharmaceutical company CEOs think about AI's ability to cure disease?
A Wall Street Journal report found that major pharma CEOs are sceptical of Silicon Valley's AI drug-discovery claims. They point to the long, complex, and heavily regulated pathway from a computational prediction to an approved therapy, suggesting a significant gap between tech industry promises and practical reality.
What is AllenAI's ToyBlocks and why does it matter for robotics?
ToyBlocks is a controlled simulated environment released by the Allen Institute for AI on Hugging Face, designed for robot learning and evaluation. It provides standardised, reproducible conditions so researchers can more reliably compare AI agents and isolate what robotic systems have genuinely learned.
