AI News Roundup: From On-Device Memory Breakthroughs to Frontier Enterprise Models

AI news this week weaves a single thread: the boundary between edge and cloud is changing, not just in theory but in architecture and practice. Apple’s AFM 3 on-device model architecture reframes the memory problem by moving the weight set out of DRAM and into NAND flash, enabling a 20-billion-parameter pool to live off the main memory. The working RAM then becomes a narrow, fast buffer that loads a fixed set of experts per prompt rather than shuttling all weights token by token. The result is a hybrid approach that keeps simpler tasks on-device and routes more demanding reasoning to AFM 3 Cloud Pro on Google Cloud, while Apple runs the on-device AFM 3 Core Advanced. Apple describes Instruction-Following Pruning (IFP) as the core mechanism: a per-prompt router loads the needed experts into RAM, and then runs the entire generation from that static expert set. Practically, this means enterprises can consider 20B parameters in flash as a new baseline for local inference, with DRAM staying as a working cache rather than the full model. The tech is still being benchmarked in a summer technical report, but the architectural shift is already a signal: the DRAM wall is moving.

Across the frontier, frontier AI moves are underway. Anthropic released two new models—Claude Fable 5 and Claude Mythos 5—aimed at broad deployment with domain-specific safeguards. Fable 5 is the default entry point for developers and apps; Mythos 5 is the higher-risk cousin reserved for trusted users under Project Glasswing, with red-teaming and a stricter access regime. Pricing sits at $10 per million input tokens and $50 per million output tokens, with discounts during rollout for subscription users. Across benchmarks and customer stories, Fable 5 demonstrates stronger capabilities in software engineering, vision tasks and long-running workflows; Stripe staff reported that a 50-million-line Ruby migration was completed in a day, a testament to extended autonomous coding. Anthropic notes more than 95% of Fable sessions rely on Fable responses alone. Yet the company also imposes a 30-day data retention policy for Mythos and Fable, with safeguards that route particularly sensitive requests to Opus 4.8. In short, Anthropic’s model tiering and governance scheme point toward enterprise automation that can do more tasks with fewer human steps, while keeping risk contained.

On the policy and market side, AI is already shaping major corporate moves and regulatory responses. Waymo’s purchase of Apple’s abandoned self-driving car test ground for about $220 million signals that AI infrastructure assets are a target for consolidation. In parallel, the AI IPO debate continues as OpenAI files to go public, highlighting profitability pressures and market dynamics that will ripple through competitors and customers alike. Regulatory attention to AI risk remains intense: the Bank of England warns about AI-generated scams as deepfakes proliferate, NHS and medical liability research flags the potential for errors in AI-assisted care, and lawmakers in the UK and US continue wrestling with safety, data governance and patching cycles. The Guardian’s coverage and other outlets frame these conversations as a broader shift toward accountable deployment rather than blanket prohibition, with industry groups seeking clear retention and data-use policies that satisfy compliance needs.

Beyond governance, the hardware and user-experience frontier is moving. Amazon’s data-center expansion with Corning fiber optics, the push to ship AI-capable infrastructure globally, and the UK’s booming yet still-maturing AI adoption landscape frame the underlying energy and latency constraints. In the consumer space, Norton’s Neo browser aims to deliver frictionless World Cup streaming by integrating anti-phishing and live streaming links directly into the browser, reducing the need to install separate apps or VPNs. The approach emphasizes calm by design and on-device privacy: personal data stays on-device unless users opt in, and the browser surfaces streaming options for a user’s market with minimal setup. In a parallel vein, China’s wind-powered underwater datacenter off Shanghai demonstrates a bold, energy-conscious approach to future-scale AI workloads, showing how infrastructure choices will shape compute availability as the AI boom intensifies.

Taken together, the week’s stories underline a central truth: the AI era is shifting from single, monolithic products to layered, governance-aware platforms that bridge devices and clouds, with enterprise architects deciding where inference should run and what should stay private. The Apple AFM 3 architecture lowers the bar for on-device intelligence, while Anthropic’s Fable 5/Mythos 5 shows how frontier capabilities can be delivered responsibly at scale. The result is not just smarter software but a new calculus for partnerships, data retention, and risk management—an AI-enabled future that must be built with both ambition and guardrails. As these threads unfold, readers should watch how disclosure, benchmarking, and policy shape what gets deployed where, and who pays for it.

1. On-device AI agents hit a hard memory limit. Apple’s new architecture routes around it
2. Crackdown on tech platforms will go ahead despite US intervention, says No 10
3. Anthropic brings Mythos to the masses with Claude Fable 5, its most powerful generally available model ever
4. Waymo Pays $220M for Apple’s Abandoned Self-Driving Car Test Ground
5. OpenAI’s IPO: Navigating Profitability and Market Dynamics
6. Palantir to sue Sadiq Khan over blocked £50m Met police contract
7. Amazon Taps Fiber Optics Producer Corning for Data Center Expansion
8. AI Adoption in the UK is Booming, but Transformation is Stalling
9. Every World Cup fan deserves a seat. Norton Neo says its free browser is the ticket
10. All signs point to Trump pushing AI growth
11. Bank of England warns of AI scams as deepfakes spread
12. Doctors and NHS could be sued for mistakes made by AI tools, report warns
13. World’s first wind-powered underwater datacentre starts operating in China