AI news for August 4 brought a mix of efficiency tricks for running models lighter and smarter, alongside efforts to plug AI agents into real-world systems.
Apple's AFM3 Narrows Down the Experts It Uses for Each Prompt
The architectural characteristics of Apple's third-generation foundation model, AFM (AFM3), are drawing attention. According to community discussions and reports, the model adopts a technique called "Instruction Following Pruning": the entire model is kept in flash storage, routing is decided just once per prompt (input), and only the small set of experts needed for that prompt is loaded into working memory.
The experts loaded per pass are on the order of 1 billion to 4 billion parameters, and they run in combination with core shared experts that stay active at all times. Rather than switching on every token, the design reportedly keeps using the same experts for the duration of a prompt.
The aim of this scheme is to save read bandwidth (memory bandwidth). Because roughly two-thirds of a model's active parameters are FFN/MLP expert weights, a 30B-active MoE, for example, can be held to the equivalent of 14B from a read-bandwidth perspective, and a 9B dense model to roughly 3B active. The activated MLP layers appear to be designed to stay within about 20% of the total.
With memory capacity and bandwidth being the biggest wall for running LLMs on device, Apple getting serious about this direction could serve as one bellwether for the spread of on-device AI.
Australia's Largest Bank NAB Moves to Test AI Agents in Production
National Australia Bank (NAB) will soon test the security and operational guardrails of an agentic AI platform, its technology chief Patrick Wright has revealed. The vision is to let AI agents interact with bank staff and systems on customers' behalf, part of a broader move by the country's major banks to deploy the technology more widely.
Wright, however, points out that designing the boundary of what agents can do — and how far — remains a challenge. In a highly sensitive domain like finance, deployment without guardrails is hard to pull off.
AI agent reliability is an industry-wide theme. There have even been reported cases of agents confirming orders that supposedly didn't exist, and the question being asked now is less about raw performance than about how much can be delegated in real-world operation.
AI Agent Market: Analysis Finds Over 15,000 Submissions in the First Half of 2026
According to a published analysis, more than 15,000 agent submissions were recorded in the AI agent market in the first half of 2026. Titled "AI Agent Ecosystem Trends 2026 H1," the report aims to take a bird's-eye view of the agent ecosystem, surfacing both the market's rapid expansion and the reality that competition is already getting crowded.
Vast, an Alibaba-Backed 3D Modeling Startup, Weighs a Hong Kong IPO
According to reports, Vast, a 3D modeling startup backed by Alibaba, is considering an initial public offering (IPO) in Hong Kong. The 3D generation and modeling field has drawn strong investor interest as a generative AI application, and fundraising activity there continues.
Shionogi Lifts Generative AI Accuracy from 50% to 90%
On the Japanese front, Shionogi is pushing generative AI adoption forward in internal operations. In the pharmaceutical industry — often considered difficult for adoption because of data confidentiality and sheer data volume — the company reportedly raised the accuracy of generative AI from 50% to 90%. Details on the methods are limited, but the fact that it tuned generative AI to "production-usable accuracy" in an industry that handles confidential data should offer a useful reference for sectors facing similar challenges.